Semantic Segmentation Project

Semantic Segmentation for the Describable Textures Dataset (DTD)

Download DTD Dataset

To download the DTD Dataset, you can use the utility files dtd_loader_color_patches.py and dtd_loader_main.py, patches will split the images in multiple patches and main will download the entire image.

Both can be called using, within the files the data directory can be specified, for the location of the dataset:
python -m src.utils.dtd_loader_main

Training

To train a model, you can run:
python -m src.train

All of the configurations for the training can be done in the settings.py script.

Currently these models are implemented:

• Simple FCN (3 convs)
• U-Net
• ResNet (with transposed convs as decoder)
• ResNeSt (with transposed convs as decoder)

Inference

To run inference on a model, which will show the segmentation mask of a random sample in the validation set, you can run:
python -m src.inference